Research On Unstructured Persona-Oriented Dialogue Generation With Neural Topical Expansion

With the development of artificial intelligence,more and more scholars and experts pay attention to the dialogue system which is one of the core tasks of natural language processing.In recent years,the boom in intelligent products such as machine intelligence customer service and chatbots also reflects people's demand for intelligent dialogue systems in real life.Dialogue systems can be divided into task-oriented dialogue systems for the purpose of completing specific tasks and non-task-oriented dialogue systems for providing chat services(also called open-domain dialogue system).In order to make machines like people in the chat to maintain persona consistent,persona-oriented dialogue systems appeared.Persona information can be divided into structured persona information represented by key-value pairs and unstructured persona information described by natural sentences,and unstructured persona information is more difficult to understand and process.Unstructured Persona-oriented Dialogue Systems has been demonstrated effective in generating persona consistent responses by utilizing predefined natural language user persona descriptions(e.g.,"I am a vegan").However,the predefined user persona descriptions are usually short and limited to only a few descriptive words,which makes it hard to correlate them with the dialogues.As a result,existing methods either fail to merge the persona description into response beacuse of or use them improperly when generating persona consistent responses.To address the above problem,this article analyzes the thinking process of people getting high-quality responses and proposes a neural topical expansion framework,namely Persona Exploration and Exploitation(PEE),which is able to extend the predefined user persona description with semantically correlated content and utilize predefined and extended persona information to generate dialogue responses.PEE consists of two main modules:persona exploration and persona exploitation.Persona exploration learns to extend the predefined user persona description by mining and correlating with existing dialogue corpus using a variational auto-encoder(VAE)based topic model.Persona exploitation learns to generate persona consistent responses by utilizing the predefined and extended user persona description.There are three components of persona exploitation stage:multi-source sequence encoder,persona information retrieval and personaoriented response decoder.In addition,in order to make persona exploitation learn to utilize user persona description more properly,this paper also introduces two persona-oriented loss functions:Persona-oriented Matching(P-Match)loss and Persona-oriented Bag-of-Words(P-BoWs)loss which respectively supervise persona selection in persona information retrieval and decoder.This article conducts extensive experiments on the dataset Persona-Chat and introduces an additional dataset for the topic model our approach PEE outperforms state-of-the-art baselines,in terms of both automatic evaluations and human evaluations.In ablation experiment,this paper analyzes the effects of persona exploration and two persona-oriented loss functions in the model.In addition,this paper conducts experiment of the persona use ratio and example analysis,especially verify the effect of mutual-reinforcement multi-hop memory retrieval.